Multi-agent consultation for architecture decisions. Dispatches 3 specialized reviewers in parallel (contrarian, user advocate, meta-process) to challenge a plan or ADR before implementation begins, producing a synthesis with a PROCEED or BLOCKED verdict. Use for "consult on this ADR", "challenge this design", "review before implementing", "should we proceed", or any Medium+ architecture decision. Do NOT use for trivial changes, simple bug fixes, or decisions already in implementation.

A/B test agent variants measuring quality and total session token cost across simple and complex benchmarks. Use when creating compact agent versions, validating agent changes, comparing internal vs external agents, or deciding between variants for production. Use for "compare agents", "A/B test", "benchmark agents", or "test agent efficiency". Do NOT use for evaluating single agents, testing skills, or optimizing prompts without variant comparison.

Wabi-sabi-aware 4-phase article evaluation: Fetch, Validate, Analyze, Report. Use when user wants to evaluate an article for voice authenticity, check voice quality, review article voice patterns, or validate content against a voice profile. Use for "evaluate article", "check voice", "is this authentic", "review my article", or "voice evaluation". Do NOT use for writing articles, editing content, or generating voice profiles without an existing article to evaluate.

Safe bulk editing across multiple Hugo markdown posts: find/replace, frontmatter updates, content transforms with mandatory preview before apply. Use when user needs batch text replacement, bulk frontmatter field changes, heading/link/whitespace normalization, or regex-based content transforms across posts. Use for "batch edit", "find and replace across files", "add field to all posts", "bulk update tags". Do NOT use for single-file edits, structural refactoring, or content generation.

Read public Bluesky feeds via the AT Protocol API. Fetch recent posts from any public profile or search a profile's posts by keyword. No authentication required. Use for "read bluesky", "check bluesky feed", "search bluesky posts".

Generate and validate Git branch names from commit messages or descriptions. Use when creating branches, generating names for /pr-sync, validating existing branch names, or converting conventional commits to branch prefixes. Triggers: "branch name", "create branch", "name this branch", "validate branch". Do NOT use for git operations (checkout, merge, delete), branching strategies, or branch protection rules.

Compose valid pipeline chains from the step menu for each subdomain in a Component Manifest. Validates type compatibility using artifact-utils.py. Produces Pipeline Spec JSON as the intermediate representation consumed by pipeline-scaffolder. Use when domain-research has completed and produced a Component Manifest with subdomains and task types. Do NOT use for scaffolding (that is pipeline-scaffolder), domain discovery (that is domain-research), or modifying existing pipelines.

Cobalt Core infrastructure knowledge: KVM exporters, hypervisor tooling, OpenStack compute.

Systematic detection and prioritization of neglected code quality issues: stale TODOs, unused imports, deprecated functions, high complexity, dead code. Use when user requests "code cleanup", "find TODOs", "technical debt scan", or "quality of life fixes". Do NOT use for bug fixing (use systematic-debugging), feature work (use test-driven-development), or formatting-only (use code-linting).

Run Python (ruff) and JavaScript (Biome) linting, formatting, and code quality checks with auto-fix support. Use when code needs linting, formatting, or style checking before commits. Use for "lint", "format", "ruff", "biome", "code style", or "check quality". Do NOT use for comprehensive code review (use systematic-code-review).

Systematic 4-phase codebase exploration: Detect, Explore, Map, Summarize. Use when starting work on an unfamiliar codebase, onboarding to a new project, reviewing a repository for the first time, or building context before debugging or code review. Use for "explore codebase", "what does this project do", "understand architecture", or "onboard me". Do NOT use for modifying files, running applications, performance optimization, or deep domain analysis.

Review and fix comments containing temporal references, development-activity language, or relative comparisons. Use when reviewing code comments, preparing documentation for release, or auditing inline comments for timelessness. Use for "check comments", "temporal language", "comment review", or "fix docs". Do NOT use for writing new documentation, API reference generation, or code style linting unrelated to comment content.

Unified 3-wave code review: Wave 0 auto-discovers packages/modules and dispatches one language-specialist agent per package for deep per-package analysis. Wave 1 dispatches 11 foundation reviewers in parallel (with Wave 0 context). Wave 2 dispatches 10 deep-dive reviewers that receive Wave 0+1 findings as context for targeted analysis. Aggregates all findings by severity, then auto-fixes ALL issues. Covers per-package deep review, security, business logic, architecture, error handling, test coverage, type design, code quality, comment analysis, language idioms, docs validation, performance, concurrency, API contracts, dependencies, error messages, dead code, naming, observability, config safety, and migration safety. Use for "comprehensive review", "full review", "review everything", "review and fix", or "thorough code review". Do NOT use for single-concern reviews (use individual agents instead).

Create a new pipeline from a task description. Fans out agent, skill, and hook scaffolding in parallel, then integrates into the routing system.

Create a new voice profile from writing samples. 7-phase pipeline: Collect, Extract, Pattern, Rule, Generate, Validate, Iterate. Wabi-sabi (natural imperfections as features) is the core principle. Use when creating a new voice, starting voice calibration, or building a voice profile from scratch. Use for "create voice", "new voice", "build voice", "voice from samples", "calibrate voice". Do NOT use for generating content in an existing voice (use voice-orchestrator), editing content (use anti-ai-editor), or comparing voices (use voice-calibrator compare mode).